☁️BizyAir Lying Sigma Sampler

The name gives away the trick, and the node's own tooltip is refreshingly direct about it: this wraps another sampler and lies to the model about how much noise (sigma) is actually present at each step. dishonesty_factor at its default of -0.05 means the sampler tells the model the sigma is 5% lower than reality, so the model believes it's further along in denoising than it really is. Push it more negative and the lie gets bigger.

Why you'd want to lie to your own model

Diffusion models make different decisions depending on how much noise they think remains. Tell one it's nearly done (a lower sigma than the truth) and it tends to commit to sharper, more confident detail earlier, since a model that thinks noise is almost gone stops hedging and starts finishing. That's the whole appeal here: it's a cheap way to push detail and sharpness without changing steps, cfg, or the base sampler itself, you're just feeding the existing sampler a slightly falsified read of where it's at.

It's not free, though. Lie too aggressively (too negative a dishonesty_factor) and you'll start seeing the same kind of artifacts you'd get from genuinely rushing the denoising process, over-baked detail, texture that looks forced rather than resolved. The default -0.05 is a conservative starting point for a reason.

The inputs that matter

sampler is a required SAMPLER input, this node wraps something else rather than being a complete sampler on its own, so you'll typically feed it the output of BizyAir_KSamplerSelect or a similar sampler-producing node upstream. dishonesty_factor is the core control, negative values reduce the sigma the model sees (the "lying nearly done" trick above), and the field has no upper cap so you can push it positive too, exaggerating remaining noise instead, though that's the less common use.

start_percent and end_percent (0.1 and 0.9 by default) window where in the sampling process the lie actually applies, it doesn't run for the very first or very last 10% of steps by default. That matters because lying about noise level at the very start (before the model has established composition) or the very end (when there's barely any noise left to misrepresent) tends to do more harm than good, keeping the effect to the middle stretch of sampling is where it earns its keep.

Output is a SAMPLER, wrapped and ready to feed into a custom-sampling chain's actual denoise-loop node further downstream.

Where it fits

This is a custom-sampling node, not a drop-in replacement for the regular KSampler's sampler_name dropdown. You build it into a chain: pick a base sampler with BizyAir_KSamplerSelect, wrap it with this node, and feed the result into whatever node in your graph actually runs the denoise loop against a BIZYAIR_MODEL.

Installing it

Through ComfyUI Manager: search "BizyAir", install, restart. Or clone it manually:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

You'll need a BizyAir/SiliconFlow API key set up before anything downstream of this node executes, via the "click to login" prompt on first use. Signup goes through phone verification, and it's worth knowing ahead of time that some users outside China have hit friction there.

Common issues

Overcooked, artifact-heavy output is almost always dishonesty_factor set too aggressively, back it toward zero (or toward the -0.05 default) before assuming the base sampler or model is at fault. If the effect seems to do nothing at all, double-check start_percent/end_percent haven't been narrowed down to nothing, and confirm the sampler input is actually coming from a real sampler-producing node rather than left unconnected.

As with every node in this pack, if the whole chain stalls or errors with no clear workflow cause, the API key is the first thing to check. And separately: if BizyAir shows up in your custom_nodes folder without you having installed it, that's a known ComfyUI Manager quirk (misresolving an unrelated missing-node error on an older shared workflow), not malware, and it's safe to remove through the Manager.